AI Agents Fail in Novel Ways, Put Businesses at Risk
Microsoft researchers identify 10 new potential pitfalls for companies who are developing or deploying agentic AI systems, with failures potentially leading to the AI becoming a malicious insider.
Research examines attack methods, defenses, and vulnerabilities, helping security teams understand risks and improve protection.
Search across headline titles and summaries.
Background for this topic.
Research is the systematic study of technologies, systems, attack methods, vulnerabilities, and defensive techniques to establish evidence and produce new findings. In information security, it includes work such as discovering flaws in software or protocols, analyzing malware and attacker behavior, testing cryptographic designs, and evaluating security controls. News under this tag may describe a proof of concept, a measurement study, or a proposed technique rather than a confirmed real-world attack.
For practitioners, research can change how risks are prioritized and mitigated. A demonstrated vulnerability may require vulnerability-management teams to verify affected assets, apply fixes, or add compensating controls; responsible disclosure gives developers time to assess and remediate before technical details enable exploitation. Research involving live systems, personal data, or offensive tooling also raises privacy, authorization, dual-use, and ethical concerns. Sound findings should state their assumptions, scope, limitations, and reproducibility, since laboratory results do not automatically show that an attack is practical in every environment.
Microsoft researchers identify 10 new potential pitfalls for companies who are developing or deploying agentic AI systems, with failures potentially leading to the AI becoming a malicious insider.
Researchers from Aon's Stroz Friedberg incident response firm discovered a new attack type, known as "Bring Your Own Installer," targeting misconfigured SentinelOne EDR installs.
Researchers from Arctic Wolf Labs detailed a new spear-phishing campaign that targets hiring managers and recruiters by posing as a job seeker.